Automatic detection of contouring errors using convolutional neural networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiation Oncology

Background:

• Accurate delineation of normal structures in head and neck imaging is crucial for radiation therapy planning.
• Existing autocontouring tools require validation and error detection mechanisms.
• Manual contour verification is time-consuming and prone to inter-observer variability.

Purpose of the Study:

• To develop an automated tool using convolutional neural networks (CNNs) for head and neck normal structure autocontouring.
• To enable automatic detection of contouring errors generated by a clinically validated autocontouring tool.
• To enhance the efficiency and accuracy of contour quality assurance in radiation therapy.

Main Methods:

• A CNN-based autocontouring tool was developed for 16 head and neck structures.
• The tool was trained and validated using computed tomography (CT) scans and clinical contours from 3495 patients.
• Accuracy was assessed using Sørensen-Dice similarity coefficients (DSC) and Hausdorff distances; error detection capabilities were evaluated on patient datasets.

Main Results:

• The CNN tool achieved high average DSC (e.g., 98.4% for brain, 89.1% for eyes) and low Hausdorff distances.
• The error detection tool correctly identified clinically unacceptable contours with high proportions (0.99/0.80).
• The tool demonstrated robust performance in identifying minor and major contouring errors from a multiatlas-based autocontouring system (MACS).

Conclusions:

• The developed CNN-based autocontouring tool demonstrates strong performance on diverse datasets.
• CNN algorithms can effectively identify suboptimal contours from validated autocontouring systems.
• This tool offers a promising solution for automatic verification of multiatlas-based autocontouring system contours, improving quality assurance.